Vector retrieval beats hybrid RAG in economic bulletin drafting

Giuseppe Bruno compares vector, graph and hybrid retrieval-augmented generation (RAG) for controlled Economic Bulletin drafting.

On a 0–100 scale, average overall quality scores were 95.36 for vector, 91.67 for hybrid and 86.68 for graph retrieval.

Faithfulness scores were 98.18, 100.00 and 99.09, respectively; traceability scores were 97.62, 94.32 and 91.70. The planner selected an average of 19.64 evidence items for vector retrieval and 24.00 each for hybrid and graph retrieval.

Under the commonly used 24-item selection cap, graph records competed with richer vector passages.

The paper attributes hybrid's lower overall score to this displacement and metrics that favor direct textual support, rather than demonstrating an inherent weakness of hybrid retrieval.

The pipeline combines ChromaDB semantic retrieval with a Neo4j knowledge graph.

Initially, graph queries returned empty placeholders for 9 of 11 chapters because assumed metadata fields did not match the database schema.

A schema-aware fallback restored substantive evidence, and all graph runs passed verification.

Separate stages collect and normalize evidence, plan sections, draft deterministically, verify citations and support, and assemble LaTeX output.

Optional local-model refinement preserves evidence markers; the study reports the author's views, not the Bank of Italy's position.